In Spring 2017, the Wharton Analytics Fellows were tasked with helping a MLB baseball team answer the following question: how can we predict and optimize the number of tickets sold across our various ticket categories (group tickets, special event tickets, etc.)? To answer this question, we tested a variety of different models and features before building a predictive engine in R that could intake user inputs and forecast ticket sales over time. As a result of this engagement, our client was able to:

• Predict ticket sales across three major ticket categories
• Model results over time to drive decision-making
• Mine for insights throughout the modeling process:
  • Do specific features, such as our team's win record, have a significant impact on ticket sales?
  • How do ticket sales vary over time?
  • How are ticket sales concentrated for each group?
  • At what point in time do group sales drop-off?

Specific names, data sources, and insights have been withheld to preserve our client's privacy.

• Frame business issue and explore available data
• Load data sources
  • Join tables
  • Set parameters
  • Rename columns
• Engineer features
  • Scrape external data sources (Baseball-Reference.com) to develop additional features (e.g., win-loss difference, home-span, etc.)
  • Explore different feature combinations using our 50+ variables
• Clean and filter dataset before adding new features
  • Remove NAs
  • Limit timeframe
• Build statistical models and visualize descriptive statistics
  • Select features for use in modeling
  • Divide into test and training sets (80/20%)
  • Models tested
    • Multivariate linear regression
    • Random Forest
    • NBD regression
  • Perform 10-fold cross-validation to optimize parameters
  • Visualize results
• Compare error terms and visuals to identify best model and covariates
  • Minimize out-of-sample forecast error (preferring Median Average Percent Error to other potential error metrics)
  • Examine predicted vs. actual plots to visually inspect fit
  • Review forecasts to ensure they align with non-time-varying estimates
  • Select final model (NBD)
    • Though the boosted Random Forest’s error terms were slightly better, we’d prefer not to use a “black box” model due to its uninterpretable results
    • The NBD regression is relatively simple to implement; many different tools and technologies can automatically run this type of model (e.g., Excel)
    • Our selected model and features are well-aligned with academic research within this space (Fader, Suher, etc.)
    • There wasn’t much variation in percent error between model types across our three ticket categories
• Develop user-friendly engine (called "Final Model")
  • Automate formulas
  • Remove excess steps / features
  • Intake user inputs
  • Export forecasts

There are two R files in this repository:

• "Model Selection" - Test OOS MAPE of multiple models before choosing the best fit
• "Final Model" - Intake user inputs via .csv before making final forecasts using the NBD regression

• Scale ticket sales to limit effects of outliers and improve computational performance
• Test more sophisticated models (e.g., RNNs)
• Trim code for efficiency